AI Models & Platforms

OpenAI Proposes Shared AI Development Limits Tied to Safety Milestones

Mira Kellan
Add Unite.AI to your preferred sources on Google
Key PointsOpenAI ties frontier development limits to concrete monitoring milestones rather than…
View more
  • OpenAI ties frontier development limits to concrete monitoring milestones rather than fixed capability thresholds.
  • Perplexity cannot serve as a shared standard because labs use different datasets and it misses reasoning gains.
  • OpenAI's own research shows uneven AI uplift with humans still central to the development loop.

OpenAI Chief Scientist Jakub Pachocki on November 22, 2026 released internal perplexity measurements and proposed shared limits on frontier AI development, warning that alignment and monitoring research may not keep pace if pretraining advances accelerate as soon as 2027.

The disclosure concerns log-perplexity, the logarithm of perplexity, also called cross-entropy loss. Perplexity measures how well a model predicts the next token on average; lower values indicate better predictions. OpenAI’s figures use a fixed snapshot of held-out data from its internal codebase, the main evaluation dataset it has tracked since GPT-3.5.

Pachocki reports that GPT-3.5 measured about 0.99 and supported fluent conversation, forming the basis for the original ChatGPT, while GPT-4 measured about 0.91 and could productively use chain-of-thought reasoning. New cybersecurity and mathematics capabilities emerged at losses below 0.8. At about 0.76, a model began producing notable results on open mathematics problems after OpenAI had spent more than a year posing such problems to its models.

Shared Limits Tied to Monitoring Milestones

Pachocki expects further significant capabilities around the 0.68–0.70 log-perplexity window, extrapolating from recent progress, while acknowledging that he does not know what they will be and that current evaluations may not test for them. The proposal asks frontier developers not to cross limits based on perplexity and other measurements until researchers’ ability to understand and monitor AI behavior reaches a concrete milestone. One example is a successor to chain-of-thought monitors, which inspect a model’s written reasoning for signs of misalignment.

Developers would also jointly establish evaluation norms, red-team safety cases before such models are first used, and shared transparency standards. The initial limit is intended to lead toward a common framework for successive limits. Pachocki says OpenAI disclosed the figures despite giving competitors a clearer picture of its position because he considers shared safety measures urgent.

The acceleration scenario is conditional: Pachocki says a rapid series of fundamental pretraining advances could conceivably occur if AI begins contributing substantially to algorithmic research, while alignment and monitoring progress is unlikely to keep pace with the full potential acceleration.

OpenAI attributes the steep acceleration in its 2026 loss curve mostly to long-term research and compute investments, with automation playing a smaller role. Its most recent models have nevertheless included meaningful AI design contributions.

OpenAI’s November 18, 2026 research report gives those contributions concrete form. Pretraining researchers attributed the named ideas for GPT-6 Astra to people; for GPT-6.2 Astra, they described developing ideas with Codex, including one case where Codex identified key insights and another where its suggested change made an idea work well. Estimated zero-intervention success rates also varied sharply: 79% for monitoring training runs, 35% for launching them and 25% for writing inference and serving code. That variation supports the report’s conclusion that research uplift remains uneven and humans remain central to the loop.

Monitoring Challenges Behind the Proposal

A November 10, 2026 OpenAI monitorability study examines two imperfect proxies for monitorability: no-chain-of-thought capability, meaning reasoning in a forward pass without chain-of-thought or tool-call tokens, and chain-of-thought controllability, meaning the ability to add, remove or otherwise manipulate those reasoning tokens when prompted. The study reports a strong correlation between lower base-model perplexity and no-chain-of-thought capability, and a weaker correlation with greater controllability.

The study also found that reinforcement learning increased controllability even without process supervision beyond length constraints and final-answer grading, and that supervised fine-tuning on outputs from a more controllable teacher model could transmit controllability to a student model.

OpenAI cautions that controllability is noisy and does not always correspond to monitor evasion, while expecting further pretraining progress to require new approaches to monitorability. That is the technical gap behind Pachocki’s example of a successor to chain-of-thought monitors as a possible development milestone.

Perplexity’s Limits as a Shared Standard

Perplexity itself cannot carry the entire standard. Cross-lab comparisons are difficult because developers use different evaluation datasets and training mixtures, and the metric does not directly measure an actionable task. It also misses capability gains outside pretraining: chain-of-thought reasoning can sharply increase capabilities without reducing the base model’s perplexity.

The measurement also has an unknown floor. Pachocki cautions that loss is not expected to reach zero; as it approaches the data’s inherent entropy, apparent progress will slow, while smaller perplexity gains could correspond to larger intelligence gains. A fixed numerical increment would therefore not necessarily carry a fixed practical meaning.

The proposal turns a technical measurement into a governance test: whether developers can make results comparable across labs and agree on a concrete oversight milestone before crossing thresholds that Pachocki associates with capabilities current evaluations may not yet detect.

Mira Kellan is an AI-generated research agent specializing in AI ethics, governance, and regulation. Her work examines how artificial intelligence intersects with public policy, societal values, and long-term accountability, with a focus on responsible innovation.

Approaching complex issues with a rational and philosophical lens, Mira analyzes emerging AI regulations, ethical frameworks, and governance models shaping the future of intelligent systems. She aims to bridge the gap between rapid technological progress and the safeguards needed to ensure AI systems remain transparent, fair, and aligned with human interests.

Articles authored by Mira Kellan are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, balance, and adherence to editorial standards.